WIP: Convert zero-inflated distributions to Mixture subclasses

[fonnesbeck]

Replaces #2231, which I messed up.

[AustinRochford]

The test_scalar_shape errors would seem to be due to the use of size here.

[AustinRochford]

I will continue working through the other sampling errors when I get the chance this week; the nan for the total probability of ZeroInflatedBinomial is still puzzling to me.

[AustinRochford]

I will continue working through the other sampling errors when I get the chance this week; the nan for the total probability of ZeroInflatedBinomial is still puzzling to me.

[fonnesbeck]

If it doesnt work out we can always pull the plug and go back to the zero-inflated subclasses of Discrete. Less elegant, but they passed the tests with flying colors.

[AustinRochford]

@fonnesbeck yes, I was thinking that it makes sense to put ZIB in the old way. I understand the problem now and think I can refactor Mixture to make it easier to change in the future, but it will take a while

[fonnesbeck]

OK, I will change this to a WIP and create a PR with the working zero-inflated classes. We can target this for post-3.1.

[twiecki]

Closing because of age. Feel free to reopen if you want to fix this up.